AI tools like ChatGPT and Claude are everywhere now, and a lot of people are using them to ask questions about their health. I’m seeing this regularly now in my practice – clients showing up with multi-page protocols, supplement stacks, and detailed explanations of what’s “really going on” with their labs.
I’m not here to tell you to stop doing this. You’re going to do it anyway, and honestly, there’s value in these tools when used well. What I want to do is help you use them in ways that are actually useful rather than ways that generate noise, waste money on unnecessary supplements, and frustrate whatever clinician you’re working with.
What AI is Actually Good At
AI tools are helpful for organizing information and generating questions. If you’ve got a stack of lab results and you’re not sure what to ask your doctor about, running them through ChatGPT can help you articulate what you’re confused about. If you’re trying to understand a diagnosis, AI can explain concepts in plain language and help you figure out what you don’t understand.
They’re also decent at summarizing long reports or research papers. If your doctor handed you a 15-page radiology report and you want the gist, AI can help with that.
What AI is Bad At (And Why)
The problem is AI tools are trained on enormous amounts of text from the internet, and a lot of that text is wellness marketing, supplement company content, and health influencer material. The training data doesn’t distinguish between a well-designed randomized controlled trial and a blog post about how turmeric cured someone’s aunt’s inflammation.
More importantly, these systems are optimized to be helpful. They want to give you answers. They want to give you actionable recommendations. They have no incentive to say “I don’t know” or “you probably don’t need to do anything about this” or “that lab value is technically out of range but clinically meaningless.”
The result is supplement sprawl. You mention three symptoms and a few lab values, and suddenly you’re looking at a protocol with 12 different supplements, each one “supported by research” in some technical sense – maybe a cell study, maybe a rat study, maybe one small trial in a population nothing like you.
AI also can’t see your full picture. It doesn’t know your complete medication list, your other diagnoses, how you actually feel day to day, or whether that “functionally low” B12 is actually a problem or just where your body sits normally. It can’t examine you. It can’t ask follow-up questions the way a clinician can. It’s working with whatever fragments you gave it and filling in the gaps with statistical patterns from its training data.
How to Get Better Results
If you’re going to use AI for health questions – and again, you probably are – here’s how to make the output actually useful.
Give it real data, not summaries. Don’t tell the AI “my B12 was low.” There’s a huge difference between a B12 of 180 and a B12 of 350 that some functional medicine reference range calls “suboptimal.” Paste in the actual values with units, reference ranges, and dates. If you don’t have the numbers, the AI is just guessing, and guessing confidently.
Share full conversations, not fragments. If you want a clinician to review what an AI told you, don’t copy-paste just the final recommendations. Share the whole conversation including your prompts. Context matters – both for understanding what the AI was responding to and for seeing where it might have gone off track. Better yet, share the conversation link with your clinician.
Ask it to critique itself. This is the most important thing, and I’ve included a prompt below that you can paste into any AI health conversation. Before you take AI-generated advice to your clinician (or worse, start acting on it yourself), ask the AI to re-read the entire conversation and evaluate its own recommendations for evidence quality, safety concerns, and overreach.
This works surprisingly well. These systems will acknowledge when they’re extrapolating from weak data if you explicitly ask them to. They’ll flag interactions they glossed over. They’ll admit that the “research” behind something is mostly cell studies. They just don’t do this automatically because they’re optimized to give you confident, helpful answers.
What Your Clinician Will (and Won’t) Do With This (Maybe)
Speaking from my own practice, when clients bring me AI-generated protocols, I treat them as brainstorming material and a window into how someone is thinking about their health. I’ll tell you which pieces look reasonable, which are low-priority, which might actually cause problems, and which are solving issues you don’t have.
What I won’t do is manage or supervise you taking a 15-supplement stack designed by ChatGPT. I won’t replace your existing medications with AI suggestions. And I’m not going to debate every line of a protocol – I’ll pull out what actually matters and set aside the rest.
Most clinicians worth working with will have a similar approach. They’ll engage with your questions and curiosity, but they’re not going to rubber-stamp a protocol just because an AI generated it.
The Self-Critique Prompt
Before sharing any AI health conversation with a clinician – or before acting on it yourself – paste the following into the chat and let the AI review its own work:
I’m going to share this conversation with my clinician. Before I do, I want you to re-read EVERYTHING above (my questions and your answers) and then do a critical self-review using the following rules:
- Evidence grading
For every supplement, herb, lab interpretation, or intervention you recommended, briefly rate the strength of evidence:
- STRONG: multiple human randomized controlled trials or high-quality systematic reviews showing meaningful benefit for my specific condition or a very close analog.
- MODERATE: some human data (small trials, observational studies) with plausible benefit but limitations or mixed results.
- PRELIMINARY: limited human data, indirect evidence, or studies in related conditions only.
- SPECULATIVE: mostly animal studies, cell studies, mechanistic reasoning, or marketing claims with little or no human data.
Label each recommendation with one of these four categories.
- Scope and overreach
Point out where you inferred more than the data supports – for example, treating “lab slightly out of range” the same as “severely abnormal,” or assuming causality where there’s only correlation. Identify places where you assumed my situation is simple when it might be complex.
- Lab interpretation hygiene
If I did NOT give you exact lab values with units, reference ranges, and dates, state clearly that your lab comments are low-reliability guesswork. Distinguish between “low by standard lab range” and “low only by functional or tightened ranges.”
- Safety and stacking
Flag any recommendations that could interact with common medications (diabetes, blood pressure, anticoagulation, seizures, psychiatric medications) or increase bleeding risk, electrolyte imbalance, liver stress, or sedation. If you suggested more than 5-7 new supplements at once, recommend a smaller starting set.
- Hype vs. reality
For heavily marketed supplements (turmeric/curcumin, resveratrol, NAD boosters, adaptogens, etc.), state clearly whether the evidence for my conditions is mainly preclinical (cells/animals) or human clinical with meaningful effect sizes. If human trials have been disappointing or inconsistent, say so.
- Uncertainty summary
List what you’re MOST confident about and why, what you’re LEAST confident about and why, and what additional information a human clinician would need before taking any of your ideas seriously.
Return your answer as: a short summary of your overall confidence in your earlier advice, a table of recommendations with evidence grades and safety notes, and questions my clinician might want to consider.
Do NOT generate new protocols. Just critique and organize what you already said.
Why This Matters
The goal here isn’t to discourage you from being curious about your health or from using tools that are freely available. The goal is to help you use those tools in ways that generate signal rather than noise.
A good AI-assisted question is one that helps you have a better conversation with your clinician. A bad one is a 47-item supplement protocol based on vague symptom descriptions and “low” labs without numbers. The difference often comes down to how you use the tool and whether you ask it to check its own work before you act on anything.
Your clinician has something AI doesn’t: the ability to integrate information about your actual body, your actual life, and the actual clinical significance of your labs. Use AI to prepare for that conversation. Don’t use it to replace it.
-Thomas